Figure 2:

Study schematic. Image volumes were compressed by using either the octree approach or conventional downsampling. Each volume was then used to train a semantic segmentation convolutional neural network. In addition to losing boundary information in the image during downsampling, spatial downsampling leads to postsegmentation upsampling, which can lead to errors in segmentation. The octree representation can avoid these limitations using an intensity tolerance, τ. Res. = resolution.